Dual-Path Region-Guided Attention Network for Ground Reaction Force and Moment Regression

TL;DR

This work tackles the challenge of estimating three-dimensional ground reaction forces and moments from high-density insole plantar pressure data, aiming to enable ambulatory gait analysis outside the lab. It introduces DP-RGNet, a dual-path architecture that combines anatomy-informed region-guided attention with a global context pathway, augmented by dynamic Center of Pressure encodings and temporal modeling via BiLSTMs. The model demonstrates strong performance gains over CNN and CNN+LSTM baselines on a custom insole dataset and a public pressure-mat dataset, achieving a six-component NRMSE of 5.78% on the insole data and 1.42% for vertical GRF on the public data. These results highlight the value of integrating biomechanical priors with data-driven learning for robust, real-world GRF/GRM estimation from wearable sensors, with implications for clinical gait assessment and rehabilitation monitoring.

Abstract

Accurate estimation of three-dimensional ground reaction forces and moments (GRFs/GRMs) is crucial for both biomechanics research and clinical rehabilitation evaluation. In this study, we focus on insole-based GRF/GRM estimation and further validate our approach on a public walking dataset. We propose a Dual-Path Region-Guided Attention Network that integrates anatomy-inspired spatial priors and temporal priors into a region-level attention mechanism, while a complementary path captures context from the full sensor field. The two paths are trained jointly and their outputs are combined to produce the final GRF/GRM predictions. Conclusions: Our model outperforms strong baseline models, including CNN and CNN-LSTM architectures on two datasets, achieving the lowest six-component average NRMSE of 5.78% on the insole dataset and 1.42% for the vertical ground reaction force on the public dataset. This demonstrates robust performance for ground reaction force and moment estimation.

Figures (4)

• Figure 1: The proposed Dual-Path Region-Guided Attention Network (DP-RGNet) architecture. The model employs a shared spatial encoder feeding into two complementary pathways: the Attention Context Path (Path A) and the Global Feature Path (Path B).
• Figure 2: Spatial Prior Definitions for Foot Segmentation (a) Schematic diagram illustrating the six functional partitions (0-5) used for spatial segmentation. (b) The explicit mapping of the high-resolution sensor grid (1024 sensors) to their corresponding partition IDs, plotted by physical (mm) coordinates.
• Figure 3: Temporal Priors for Prototype Activation The curves illustrate the mean pressure (in arbitrary units, a.u.) for each of the six functional prototypes, averaged across all subjects in the training dataset. The time axis is normalized during a complete stance phase.
• Figure 4: Measured (solid black line, with shaded gray area representing the mean $\pm$ SD) and predicted GRF/GRM components of the selected sample: DP-RGNet (red solid line), DP-RGNet (PathB-only) (red dot line), CNN-LSTM Baseline (blue solid line), and CNN Baseline (blue dot line) during the stance phase of an example step of subject 4 walking at 1.0 m/s.